Add Hybrid RAG chat with Huberman CLI example

example/answer.ts

@@ -0,0 +1,35 @@
+import { chatModel, chunkDatastore } from './tools.js';
+import type { Prompt } from '../src/prompt/index.js';
+import { Msg } from '../src/prompt/index.js';
+
+const system = Msg.system(`
+  You are a helpful and accurate Q&A bot.
+  You answer the user's QUESTION, based only on the PROVIDED_CONTEXT.
+  Your answer should be at most 100 words long.
+  You use Markdown footnotes to cite the chunkId of the PROVIDED_CONTEXT that
+  you used for each part of your answer. (Eg: [^1])
+`);
+
+function userMsg(query: string, chunks: unknown[]) {
+  return Msg.user(`
+    QUESTION: ${query}
+    PROVIDED_CONTEXT:
+    \`\`\`json
+    ${JSON.stringify(chunks, null, 2)}
+    \`\`\`
+  `);
+}
+
+export async function generateAnswer(
+  query: string,
+  history: Prompt.Msg[]
+): Promise<Prompt.Msg> {
+  const results = await chunkDatastore.query({ query, topK: 5 });
+  const chunks = results.docs.map((doc) => doc.metadata);
+  const messages = [system, ...history.slice(-5), userMsg(query, chunks)];
+  const { message } = await chatModel.run({
+    messages,
+    handleUpdate: (c) => process.stdout.write(c),
+  });
+  return message;
+}

example/cli.ts

@@ -0,0 +1,24 @@
+import 'dotenv/config.js';
+import readline from 'readline';
+import { generateAnswer } from './answer.js';
+import type { Prompt } from '../src/prompt/types.js';
+
+const history: Prompt.Msg[] = [];
+const rl = readline.createInterface({
+  input: process.stdin,
+  output: process.stdout,
+});
+
+function askQuestion(): Promise<string> {
+  return new Promise((resolve) => rl.question('> ', resolve));
+}
+
+console.log('Welcome to Huberman chat. Enter your question to begin.');
+
+while (true) {
+  const query = await askQuestion();
+  process.stdout.write('Huberman: ');
+  const answer = await generateAnswer(query, history);
+  process.stdout.write('\n\n');
+  history.push(answer);
+}